| Abstract [eng] |
3D object classification by using 2D images is used in fields, where given 3D object classes have to be recognized in any possible 2D image, by using only a subset of those images. In this master’s thesis neural capsule network modifications are compared with current state of the art solution of this problem, multi-view convoliutional network. Tested neural capsule network modifications are multi-view neural capsule networks, of which one is using view pooling layer and the other one - view capsule layer. Conducted experiments allowed to draw a conclusion that current state of the art solution is better than tested capsule network modifications. |
